foot/doc/benchmark.md

104 lines
3.8 KiB
Markdown

# Benchmarks
## vtebench
All benchmarks are done using [vtebench](https://github.com/alacritty/vtebench):
```sh
vtebench -h $(tput lines) -w $(tput cols) -b 104857600 alt-screen-random-write > ~/alt-random
vtebench -c -h $(tput lines) -w $(tput cols) -b 104857600 alt-screen-random-write > ~/alt-random-colors
vtebench -h $(tput lines) -w $(tput cols) -b 10485760 scrolling > ~/scrolling
vtebench -h $(tput lines) -w $(tput cols) -b 104857600 scrolling --fill-lines > ~/scrolling-filled-lines
vtebench -h $(tput lines) -w $(tput cols) -b 10485760 unicode-random-write > ~/unicode-random
```
They were "executed" using [benchmark.py](../scripts/benchmark.py),
which will load each file into memory, and then print it to the
terminal. This is done **20** times for each test. Then it calculates
the _mean_ and _standard deviation_ for each test.
## 2020-07-25
### System
CPU: i9-9900
RAM: 64GB
Graphics: Radeon RX 5500XT
### Terminal configuration
Geometry: 2040x1884
Font: Fantasque Sans Mono 10.00pt/23px
Scrollback: 10000 lines
### Results
| Benchmark | Foot (GCC+PGO) 1.4.2.r14 | Alacritty 0.4.3 | URxvt 9.22 | XTerm 358 |
|------------------------|-------------------------:|---------------------:|---------------:|---------------:|
| alt-random | 0.423s ±0.014 | 0.904s ±0.006 | 1.111s ±0.003 | 12.851s ±0.087 |
| alt-random-colors | 0.382s ±0.005 | 0.935s ±0.005 | 1.146s ±0.007 | 11.816s ±0.088 |
| scrolling | 1.380s ±0.048 | 1.011s ±0.012 | 1.021s ±0.016 | 38.483s ±0.122 |
| scrolling-filled-lines | 0.826s ±0.020 | 1.307s ±0.008 | 1.213s ±0.015 | 6.725s ±0.016 |
| unicode-random | 0.243s ±0.006 | 0.091s ±0.003 [^1] | 24.507s ±3.264 | 26.127s ±3.891 |
## 2020-05-31
### System
CPU: i5-8250U
RAM: 8GB
Graphics: Intel UHD Graphcis 620
### Terminal configuration
Geometry: 953x1023
Font: Dina:pixelsize=12
Scrollback=10000 lines
### Results
| Benchmark | Foot (GCC+PGO) 1.3.0.r59 | Alacritty 0.4.2 | URxvt 9.22 | St 0.8.3 | XTerm 356 |
|------------------------|-------------------------:|---------------------:|---------------:|--------------:|---------------:|
| alt-random | 0.791s ±0.080 | 1.558s ±0.038 | 1.746s ±0.065 | 2.628s ±0.085 | 1.706s ±0.064 |
| alt-random-colors | 0.830s ±0.076 | 1.587s ±0.041 | 2.049s ±0.118 | 3.033s ±0.129 | 2.109s ±0.131 |
| scrolling | 1.603s ±0.070 | 1.464s ±0.098 | 1.439s ±0.035 | 3.760s ±0.113 | 1.459s ±0.036 |
| scrolling-filled-lines | 1.888s ±0.021 | 2.334s ±0.078 | 2.145s ±0.074 | 3.372s ±0.078 | 2.144s ±0.091 |
| unicode-random | 1.545s ±0.229 | 0.164s ±0.012 [^1] | 11.180s ±0.342 | crashed | 11.389s ±0.269 |
[^1]: [Alacritty and "unicode-random"](#alacritty-and-unicode-random)
# Alacritty and "unicode-random"
Alacritty is actually **really** slow at rendering this (whether it is
fallback fonts in general, emojis, or something else, I don't know).
I believe the reason it finishes the benchmark so quickly is because
it reads from the PTY in a separate thread, into a larger receive
buffer which is then consumed by the main thread. This allows the
client program to write its output much faster since it is no longer
stalling on a blocked PTY.
This means Alacritty only needs to render a couple of frames since it
can reach the final VT state almost immediately.
On the other hand, `cat`:ing the `unicode-random` test file in an
endless loop, or just manually scrolling up after the benchmark is
done is **slow**, which besides being felt (input lag), can be seen by
setting `debug.render_timer = true` in `alacritty.yml`.